Singing Voice Conversion with Disentangled Representations of Singer and Vocal Technique Using Variational Autoencoders

Yin-Jyun Luo$^{1, 3}$, Chin-Cheng Hsu$^{2}$, Kat Agres$^{3,4}$, Dorien Herremans$^{1,3}$

$^{1}$Singapore University of Technology and Design $^{2}$University of Southern California $^{3}$Institute of High Performance Computing, A*STAR, Singapore $^{4}$Yong Siew Toh Conservatory of Music, National University of Singapore $\tt yinjyun\_luo@mymail.sutd.edu.sg$

Many-to-Many Singing Technique Conversion

We focus on singing technique conversion in this demo page.
The following are the audio samples of Fig. 2(b) in the paper.

The audio files below are all converted from Mel-spectrograms using Griffin-Lim.
Therefore, the audio "original Mel-spectrogram" are the upper bounds of the audio quality for each conversion.

We first give examples of each singing technique

Again, these samples are obtained by inverting from their Mel-spectrograms to audio.

Notice that singers could express a vocal technique distinctly, and also sing at different levels of expression at different time instants, which raises ambiguities on defining vocal techniques and poses challenges to data-driven models.